An Exploration of the Many-Objective Optimisation Process for a Class of Evolutionary Algorithms.

This empirical inquiry explores the behaviour of a particular class of evolutionary algorithms as the number of conflicting objectives to be simultaneously optimised is increased. Population-based optimisers that perform selection according to Pareto dominance and density estimation are considered. The performances of abstracted algorithms, based on decompositions of the fundamental components of a modern optimiser, are considered across a wide range of mutation and recombination operating conditions. Configuration sweet-spots for these algorithms are identified and contrasted. The classical mutation settings are shown to be a robust choice, even when the total sweet-spot is seen to contract as the number of objectives is increased. Classical settings for recombination, by contrast, are shown to work well for small numbers of objectives but lead to very poor performance as the number of objectives is increased. Mutation performance is demonstrated to be largely invariant of population size across the standard range of values. The performance of recombination can be somewhat improved by using larger population sizes. Explanations are offered for the observed behaviour of the evolutionary optimisers.